Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)



First Advisor

W. Douglas McMillin


This dissertation consists of three essays on vector autoregressive (VAR) models. The first examines lag selection criteria. Typically, in VAR models all variables have the same lag length in each equation. Hsiao (1981) and Keating (1994) suggest two ways to estimate asymmetric lag structures. In a Monte-Carlo framework, we simulate eight prespecified VAR models and observe how the lag specification methods perform. We also look at the impulse response and forecast performances of these lag selection methods. We find that the AIC criterion performs better in finite samples than the SIC criterion, and Keating's method is superior to Hsiao's method. In the second part of the dissertation, we employ the Stein-rule estimator to estimate two VAR models, one using quarterly macroeconomic data and another one using monthly macroeconomic data. The forecasts produced by VARs estimated via Stein-rule are contrasted with the forecasts produced by VARs estimated via Bayesian methods and via ordinary least squares (OLS). In general, Bayesian VARs and Stein-rule VARs produce more accurate forecasts than OLS VARs; however, Bayesian forecasts are more expensive and difficult to obtain than either Stein-rule VARs or OLS VARs. We find that usually Stein-rule VARs perform better than Bayesian VARs. The last part of the dissertation estimates VAR models with a discrete variable. The equation with the discrete variable as the dependent variable is estimated using the probit approach. The impulse response function (IRF) and variance decomposition (VDC) produced from VARs partly estimated via the probit approach are contrasted with the IRF and VDC values produced from VARs estimated via the OLS technique. The results obtained from these two alternative approaches are different. However, the results from the probit approach show that the IRF and VDC values obtained with the current technique are not plausible in terms of theoretical expectations and that the modeling and estimation of a VAR that includes a discrete variable has to be further investigated.